From 8bd37038c2b52f9d81524ceec266f043dbe8db4c Mon Sep 17 00:00:00 2001
From: lbittarello <luca.bittarello@gmail.com>
Date: Mon, 16 Mar 2026 09:37:41 +0000
Subject: [PATCH] Partially reverse 975

---
 CHANGELOG.rst                   |  6 +++++-
 src/glum/_glm.py                | 28 +++++++++++++++-------------
 tests/glm/test_glm_regressor.py | 18 ++++++++++++++++++
 3 files changed, 38 insertions(+), 14 deletions(-)

diff --git a/CHANGELOG.rst b/CHANGELOG.rst
index 0de61e4f..eadf0ecd 100644
--- a/CHANGELOG.rst
+++ b/CHANGELOG.rst
@@ -7,9 +7,13 @@
 Changelog
 =========
 
-3.2.1 - unreleased
+3.2.1 - 2026-03-16
 ------------------
 
+**Bug fix:**
+
+- Fixed an error when predicting at a specific ``alpha`` with categorical features.
+
 **Other changes:**
 
 - Downgraded log messages in ``align_df_categories`` and ``add_missing_categories`` from INFO to DEBUG, and deduplicated them so they are emitted only once per column per fitted model.
diff --git a/src/glum/_glm.py b/src/glum/_glm.py
index 6cf03366..8c9e73e9 100644
--- a/src/glum/_glm.py
+++ b/src/glum/_glm.py
@@ -889,20 +889,22 @@ def _compute_linear_predictor(
             )
 
         if alpha_index is None:
-            coef = coef_path
-            intercept = intercept_path
+            xb = X @ coef_path + intercept_path
+            if offset is not None:
+                xb += offset
+        elif np.isscalar(alpha_index):  # `None` doesn't qualify
+            xb = X @ coef_path[alpha_index] + intercept_path[alpha_index]  # type: ignore
+            if offset is not None:
+                xb += offset
         else:
-            scalar = np.isscalar(alpha_index)
-            alpha_index = np.atleast_1d(alpha_index)  # type: ignore[assignment]
-            coef = coef_path[alpha_index]  # type: ignore
-            intercept = intercept_path[alpha_index]  # type: ignore
-
-        xb = X @ coef.T + intercept
-        if offset is not None:
-            offset = np.asanyarray(offset)
-            xb += offset if xb.ndim == 1 else offset[:, np.newaxis]  # type: ignore[call-overload]
-
-        return xb.squeeze() if alpha_index is None or scalar else xb
+            _xb = []
+            for idx in alpha_index:  # type: ignore
+                _xb.append(X @ coef_path[idx] + intercept_path[idx])  # type: ignore
+            xb = np.stack(_xb, axis=1)
+            if offset is not None:
+                xb += np.asanyarray(offset)[:, np.newaxis]
+
+        return xb
 
     def predict(
         self,
diff --git a/tests/glm/test_glm_regressor.py b/tests/glm/test_glm_regressor.py
index d37a539a..351d1077 100644
--- a/tests/glm/test_glm_regressor.py
+++ b/tests/glm/test_glm_regressor.py
@@ -1202,6 +1202,24 @@ def test_predict_list(regression_data, alpha, alpha_index):
     np.testing.assert_allclose(candidate, target + 1)
 
 
+def test_predict_list_categorical():
+
+    letters = ["a", "b", "c", "d", "e", "f"]
+    rng = np.random.default_rng(42)
+
+    df = pd.DataFrame({"x": rng.choice(letters, size=100)})
+
+    df["x"] = df["x"].astype("category")
+    df["y"] = df["x"].map({v: k + 1 for k, v in enumerate(letters)})
+
+    regressor = GeneralizedLinearRegressor(alpha=[0, 2], alpha_search=True)
+    regressor = regressor.fit(df[["x"]], df["y"])
+
+    candidate = regressor.predict(df[["x"]], alpha=0)
+
+    np.testing.assert_allclose(candidate, df["y"])
+
+
 def test_predict_error(regression_data):
     X, y = regression_data